d connecting the “frontiers” of each object’s color distribution. As shown in Figure 2.1a, this approach implicitly assumes that the unknown region is a narrow band around the foreground boundary, and the skeleton of the unknown region can be represented by a chain of pixels. The model construction and alpha estimation procedure is summaried as follows: (1) Divide the chain of pixels (the skeleton of the unknown re- gion) into intervals by selected anchor points ; (2) Centered on each anchor point, de?ne a local spatial window which covers a local unknown region, and a local foreground and background region. (3) Foreground and background pixels in the local window are used to estimate a foreground and background isotropic (un- oriented) Gaussian distribution, or point mass , respectively,
全文预览
